DeepLabV < 3 + >-based tunnel face crack identification method and device and medium

Abstract

The invention belongs to the technical field of computer vision and deep learning, and particularly relates to an improved DeepLabV < 3 + > tunnel face crack recognition method and device and a medium. A tunnel face image of a construction site is obtained and preprocessed, the preprocessed tunnel face image is input into a tunnel face crack recognition model based on DeepLabV3 +, and the tunnel face crack recognition model adopts the first three layers of a ResNet series model as a backbone network of DeepLabV3 + to complete multi-scale feature extraction of the tunnel face image. The IASPP module is added behind the first layer and the third layer of the backbone network, the IASPP module adds a 3 * 3 convolutional layer in front of a 1 * 1 convolutional layer of ASPP, the aliasing effect caused by multi-level feature fusion is eliminated, and the feature quality is improved. According to the method, higher segmentation precision and clearer boundaries are obtained, and an effective technical means is provided for crack detection in underground space engineering.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, specifically to a method, device and medium for identifying cracks at the tunnel face based on DeepLabV3+, which is applicable to crack detection in underground space engineering. Background Technology

[0002] In the task of identifying cracks at tunnel face, the DeepLabV3+ model, as a classic identification model, is widely used due to its powerful feature representation capabilities and multi-level feature fusion capabilities. However, existing techniques still have some shortcomings in complex scenarios. First, cracks in tunnel face images typically occupy a small area, and background interference is significant, such as rock texture and uneven lighting, leading to insufficient accuracy in feature extraction and segmentation. Second, the DeepLabV3+ model only adds the ASPP module after the output feature map with a stride of 16 in Xception, failing to fully exploit the detailed location information contained in the output feature map with a stride of 4, resulting in an unbalanced fusion of location and semantic information. Furthermore, after multi-scale feature fusion, the ASPP module directly adjusts the number of channels through 1×1 convolutions, ignoring the aliasing effect caused by multi-level feature fusion, thus affecting feature quality. Finally, although Xception, as the backbone, uses depthwise separable convolutions to reduce computation, its network depth is relatively shallow, limiting its feature representation capabilities and making it difficult to fully extract the complex texture and detailed information of cracks in tunnel face images. These issues directly affect the accuracy and reliability of crack identification, which in turn affects the safety and construction efficiency of underground space engineering. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method, device, and medium for identifying cracks at tunnel face based on DeepLabV3+. The DeepLabV3+ model is improved to enhance its segmentation accuracy in semantic segmentation of cracks at tunnel face, providing a more effective technical means for crack detection in underground space engineering.

[0004] To solve the aforementioned technical problem, the technical solution adopted by the present invention is: a method for identifying tunnel face cracks based on DeepLabV3+, comprising the following steps:

[0005] S01. Obtain images of the tunnel face at the construction site and preprocess them;

[0006] S02. Input the preprocessed tunnel face image into the tunnel face crack recognition model based on DeepLabV3+. The tunnel face crack recognition model uses the first three layers of the ResNet series model as...

[0007] In the DeepLabV3+ backbone network, in each residual block of the second and third layers of the backbone network, the 3×3 convolution is replaced with two cascaded dilated convolutions, and the stride of the first convolution kernel of the first residual block of the second and third layers of the backbone network is set to 1; the backbone network completes the multi-scale feature extraction of tunnel face images.

[0008] The first and third layers of the backbone network are both connected to IASPP modules. The IASPP module adds a 3×3 convolutional layer before the 1×1 convolutional layer of the ASPP module to eliminate aliasing. The pooling of multi-scale features of the tunnel face image is completed through the two IASPP modules to obtain the deep feature map.

[0009] The output features of the first layer of the backbone network are processed by the IASPP module and a 1×1 convolution to obtain feature F. l1iaspp F l1conv The output features of the third layer of the backbone network are obtained by upsampling through the IASPP module to obtain feature F. l3conv Feature F l1iaspp F l1conv F l3conv After splicing, feature F is obtained. cc Feature F cc After eliminating aliasing effects and refining the features using a 3×3 convolutional layer, feature F is obtained. c33 , for F c33 Upsampling is performed, and then the result F is calculated by substituting it into the softmax function. prediction .

[0010] Furthermore, the first three layers of ResNet-101 are used as the backbone network of DeepLabV3+. The backbone network consists of an input layer, a first layer, a second layer, and a third layer connected in series.

[0011] The first layer consists of one max pooling layer and three residual blocks. Each residual block includes three concatenated 1×1 convolutional layers I, 3×3 convolutional layers, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer to the output of the three concatenated convolutional layers.

[0012] The second layer consists of four concatenated residual blocks. Each residual block includes four concatenated 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer and the output of the four concatenated convolutional layers.

[0013] The third layer consists of 23 residual blocks in series. Each residual block includes four cascaded 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer and the output of the four cascaded convolutional layers.

[0014] Furthermore, the stride of each residual block in the first layer is 1 for the 1×1 convolutional layer I, the 3×3 convolutional layer, the 1×1 convolutional layer II, and the skip connection layer.

[0015] Furthermore, in the second layer, the stride of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II in each residual block is 1, the rates of the 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4 respectively, the stride of the 1×1 convolutional layer I and the skip connection layer in the first residual block of the second layer is 2, and the stride of the 1×1 convolutional layer I and the skip connection layer in the last three residual blocks is 1.

[0016] Furthermore, in the third layer, the stride of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II in each residual block is 1, the rates of the 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4 respectively, the stride of the 1×1 convolutional layer I and the skip connection layer in the first residual block of the third layer is 2, and the stride of the 1×1 convolutional layer I and the skip connection layer in the last 22 residual blocks is 1.

[0017] Furthermore, the IASPP module includes parallel 1×1 convolutional layers, 3×3 dilated convolutional layers I, II, and III, and a global average pooling layer. The global average pooling layer is followed by a 1×1 convolutional layer and an upsampling layer. The input passes through these parallel layers to obtain the feature F. c1 F c2 F c3 F c4 F map F map The feature F is obtained after a 1×1 convolutional layer and an upsampling layer following a global average pooling layer. c5 , feature F c1 F c2 F c3 F c4 and F c5 The feature F is obtained by splicing. concat Feature F concat After passing through a series of 3×3 convolutional layers and 1×1 convolutional layers, the output of the IASPP module is obtained.

[0018] Furthermore, the rates of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 3×3 dilated convolutional layer III are 6, 12, and 18, respectively.

[0019] Furthermore, the preprocessing includes normalization, data augmentation, and label processing.

[0020] The present invention also discloses a tunnel face crack identification device based on DeepLabV3+, including a processor and a memory storing program instructions. The processor is configured to execute the tunnel face crack identification method based on DeepLabV3+ as described above when running the program instructions.

[0021] The present invention also discloses a storage medium storing program instructions, which, when executed, perform the tunnel face crack identification method based on DeepLabV3+ as described above.

[0022] The beneficial effects of this invention are as follows: This invention uses DeepLabV3+ as the base model and improves the model's feature representation ability and the semantic segmentation accuracy of tunnel face cracks by introducing an improved ResNet-101 network structure (IR-Net), an improved ASPP module (IASPP), and an improved multi-scale feature fusion module (IMSFF). First, IR-Net, as a bridge, improves the ResNet-101 network structure by using 1×1 convolutions with a stride of 2 to achieve spatial dimensionality reduction, reducing computational cost, and replacing 3×3 convolutions with two cascaded dilated convolutions to expand the receptive field. Second, the IASPP module adds a 3×3 convolutional layer before the 1×1 convolutional layer of ASPP, eliminating the aliasing effect caused by multi-level feature fusion and improving feature quality. Finally, the IMSFF module adds a layer after the output feature maps of the first and third layers of ResNet-101.

[0023] The improved model utilizes the IASPP module and stitches feature maps of different scales to more evenly integrate location and semantic information, thereby enhancing segmentation accuracy and detail capture capabilities. Experimental results demonstrate that the improved model achieves higher segmentation accuracy and clearer boundaries in the semantic segmentation of tunnel face cracks, providing an effective technical means for crack detection in underground space engineering. Attached Figure Description

[0024] Figure 1 A schematic diagram of the structure of a tunnel face crack identification model;

[0025] Figure 2 A schematic diagram of the backbone network and its corresponding residual modules;

[0026] Figure 3This is a schematic diagram of the IASPP structure;

[0027] Figure 4 This is a schematic diagram of the device described in Example 2. Detailed Implementation

[0028] Example 1

[0029] This embodiment discloses a tunnel face crack identification method based on DeepLabV3+. This method uses DeepLabV3+ as the base model and improves the feature representation ability and segmentation accuracy of the model by introducing an improved ResNet-101 network structure (IR-Net) as the backbone, as well as an improved ASPP module (IASPP) and an improved multi-scale feature fusion module (IMSFF).

[0030] Compared to Xception, ResNet-101, as a deeper network structure with stronger feature representation capabilities, can better extract the complex texture and detail information of cracks in tunnel face images. Furthermore, dilated convolutions expand the receptive field, extracting richer features without increasing computational cost. Therefore, dilated convolutions are introduced to improve ResNet-101, using the network structure up to the fourth layer as...

[0031] The DeepLabV3+ backbone replaces 3×3 convolutions with two cascaded dilated convolutions in each residual block of the second and third layers, with dilation rates of 2 and 4 respectively and a stride of 1. This achieves a receptive field size equivalent to that of four 3×3 convolution downsampling operations while maintaining the same feature map size. Simultaneously, the stride is set to 2 in the first 1×1 convolution of the first residual block in the second and third layers to achieve spatial dimensionality reduction and reduce computational cost; the stride of the 1×1 convolutions in other residual blocks is 1. The improved IR-Net network structure is as follows: Figure 2 As shown, the left figure is the overall network structure of IR-Net, and the right figure is divided into three parts from top to bottom, corresponding to the detailed structure of the first residual block in the first to third layers.

[0032] Depend on Figure 2 It can be seen that the backbone network consists of an input layer, a first layer, a second layer, and a third layer connected in series.

[0033] The first layer consists of one max pooling layer and three residual blocks in series. Each residual block includes three concatenated 1×1 convolutional layers I, 3×3 convolutional layers, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block, and the output of the skip connection layer is added to the output of the three concatenated convolutional layers to obtain the output of this residual block. The second layer consists of four residual blocks in series. Each residual block includes four concatenated 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, 1×1 convolutional layers II, and a skip connection layer. The third layer consists of 23 concatenated residual blocks. Each residual block includes four concatenated 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, and 1×1 convolutional layers II, as well as one jump connection layer. The jump connection layer connects the input and output of the residual block, and the output of the jump connection layer is added to the output of the four concatenated convolutional layers to obtain the output of the residual block.

[0034] Specifically, in the first layer, the stride of each residual block's 1×1 convolutional layer I, 3×3 convolutional layer, 1×1 convolutional layer II, and skip connection layer is 1. In the second layer, the stride of each residual block's 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II is 1. The rates of 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4, respectively. The stride of the first residual block's 1×1 convolutional layer I and skip connection layer is 2, and the stride of the last three residual blocks' 1×1 convolutional layer I and skip connection layer is 1. In the third layer, the stride of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II in each residual block is 1. The rates of the 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4, respectively. The stride of the 1×1 convolutional layer I and the skip connection layer in the first residual block of the third layer is 2, and the stride of the 1×1 convolutional layer I and the skip connection layer in the last 22 residual blocks is 1.

[0035] The ASPP module generates multi-scale feature maps through dilated convolutions at different scales, and then adjusts the number of channels using 1×1 convolutional layers. However, directly concatenating feature maps of different scales leads to aliasing, affecting feature quality. Adding a 3×3 convolutional layer before the 1×1 convolutional layer effectively eliminates this aliasing effect and improves feature quality. The improved IASPP network structure is as follows: Figure 2As shown, the IASPP module includes parallel 1×1 convolutional layers, 3×3 dilated convolutional layers I, II, and III, and a global average pooling layer. After the global average pooling layer, a 1×1 convolutional layer and an upsampling layer are connected. The input passes through these parallel layers to obtain the feature F. c1 F c2 F c3 F c4 F map F map The feature F is obtained after a 1×1 convolutional layer and an upsampling layer following a global average pooling layer. c5 , feature F c1 F c2 F c3 F c4 and F c5 The feature F is obtained by splicing. concat Feature F concat After passing through a series of 3×3 convolutional layers and 1×1 convolutional layers, the output of the IASPP module is obtained.

[0036] Specifically, the rates of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 3×3 dilated convolutional layer III are 6, 12, and 18, respectively.

[0037] To achieve multi-feature fusion, IASPP modules are added after the first and third layers of IR-Net. The 1×1 convolutions within each IASPP module adjust the number of channels in the output feature map to n. Then, the third layer is processed...

[0038] The feature map output by the IASPP module is upsampled by 4x bilinear interpolation. Simultaneously, the output feature map of the first layer is processed by a 1×1 convolution to adjust the number of feature map channels to n. The three feature maps are then fused side-by-side, followed by a 3×3 convolution to eliminate aliasing. This more evenly integrates positional and semantic information, improving segmentation accuracy and detail capture. The improved overall network structure of DeepLabV3+ after introducing IMSFF is as follows: Figure 1 As shown.

[0039] The following section details the specific implementation process of this invention from five aspects: dataset construction, data preprocessing, model implementation, model training, and model evaluation.

[0040] S101. Dataset Construction: The tunnel face image dataset used in this invention originates from the construction site of an underground engineering project. The dataset contains 1000 high-quality tunnel face images, with a uniform resolution of 512×512 pixels. Each image has been annotated by professionals, including annotations of the fracture area and background area. To ensure the diversity and representativeness of the dataset, it covers tunnel face images under different lighting conditions, different rock textures, and different fracture widths and directions.

[0041] S102. Dataset preprocessing: Before model training, the dataset is preprocessed, including the following steps:

[0042] Normalization: Normalize the pixel values ​​of the image to the range of [0,1] to speed up the convergence of the model.

[0043] Data augmentation: To improve the generalization ability of the model, data augmentation operations such as random flipping, rotation, scaling, and brightness adjustment are performed on the image.

[0044] Label processing: The labeled crack area and background area are encoded as 1 and 0 respectively to form binary labels.

[0045] S103, Model Implementation, such as Figure 1 As shown, the IR-Net module, IASPP module, and IMSFF module are constructed sequentially.

[0046] (1) The IR-Net module improves ResNet-101 by introducing dilated convolutions. The network structure before the fourth layer is used as the backbone of DeepLabV3+. In each residual block of the second and third layers, the 3×3 convolution is replaced with two cascaded dilated convolutions with dilation rates of 2 and 4, respectively, and a stride of 1. While maintaining the same feature map size, the receptive field size is equivalent to the receptive field size of four 3×3 convolution downsampling operations. Simultaneously, in the first 1×1 convolution of the first residual block in the second and third layers, the stride is set to 2 to achieve spatial dimensionality reduction and reduce computational cost. The stride of the 1×1 convolutions in other residual blocks is 1. The calculation process is as follows:

[0047] Input image Feature extraction and dimensionality reduction are performed using a 7×7 convolutional layer with a stride of 2, resulting in a 64-channel feature map.

[0048] F 1 The input is the first layer, which consists of one max-pooling layer and three residual blocks concatenated together. F 1 The feature map is calculated and reduced by using a 3×3 max pooling layer with a stride of 2 to preserve feature invariance. F mpAs input, the data passes through the first residual block, which is then concatenated with a 1×1 convolutional layer with a stride of 1, a 3×3 convolutional layer with a stride of 1, and a 1×1 convolutional layer with a stride of 1. It also includes a 1×1 convolutional layer with a stride of 1 as a skip connection. These processes sequentially calculate and output a 64-channel feature map. 64-channel feature map 256-channel feature map 256-channel feature map F l1rsb1-3 and F l1rsb1-sk The output of the first residual block is obtained by adding elements together. F l1rsb1 As input, the data is processed sequentially through the second and third residual blocks. The structures of the second and third residual blocks are identical to those of the first residual block. This process is repeated to calculate and output a 256-channel feature map. and

[0049] F l1rsb3 Input to the second layer, which consists of four cascaded residual blocks, F l1rsb3 As input, the data passes through the first residual block, which is then concatenated with a 1×1 convolutional layer with a stride of 2, a 3×3 dilated convolutional layer with a stride of 1 and a rate of 2, a 3×3 dilated convolutional layer with a stride of 1 and a rate of 4, and a 1×1 convolutional layer with a stride of 1. A 1×1 convolutional layer with a stride of 2 is also included as a skip connection. This process sequentially calculates and outputs a 128-channel feature map. 128-channel feature map 128-channel feature map 512-channel feature map 512-channel feature map F l2rsb1-4 and F l2rsb1-sk The output of the first residual block is obtained by adding elements together. F l2rsb1 As input, the data passes through the second, third, and fourth residual blocks in sequence. The structures of the second, third, and fourth residual blocks are identical to those of the first residual block. The difference lies in the stride: in the first residual block, the stride of the first 1×1 convolutional layer and the skip connection is 2, while in the second, third, and fourth residual blocks, the stride of the first 1×1 convolutional layer and the skip connection is 1. This process sequentially calculates and outputs a 512-channel feature map. and

[0050] F l2rsb4 The input is the third layer, which consists of 23 residual blocks connected in series, F l2rsb4As input, the data passes through the first residual block, which is then concatenated with a 1×1 convolutional layer with a stride of 2, a 3×3 dilated convolutional layer with a stride of 1 and a rate of 2, a 3×3 dilated convolutional layer with a stride of 1 and a rate of 4, and a 1×1 convolutional layer with a stride of 1. A 1×1 convolutional layer with a stride of 2 is also included as a skip connection. These steps sequentially calculate and output a 256-channel feature map. 256-channel feature map 256-channel feature map 1024-channel feature map 1024-channel feature map Will

[0051] F l3rsb1-4 and F l3rsb1-sk The output of the first residual block is obtained by adding elements together. F l3rsb1 As input, the data passes through the last 22 residual blocks. These last 22 residual blocks have the same structure as the first residual block, except that the stride of the first 1×1 convolutional layer and skip connections in the first residual block is 2, while the stride of the first 1×1 convolutional layer and skip connections in the last 22 residual blocks is 1. The final output is a 1024-channel feature map.

[0052] (2) The IASPP module adds a 3×3 convolutional layer before the 1×1 convolutional layer of the ASPP module to eliminate aliasing. The calculation process is as follows:

[0053] F l1rsb3 As input, the system simultaneously takes into account a parallel 1×1 convolutional layer with a stride of 1, a 3×3 dilated convolutional layer with a rate of 6, a 3×3 dilated convolutional layer with a rate of 12, a 3×3 dilated convolutional layer with a rate of 18, and a global average pooling layer. The parallel computation outputs the following layers: F map The input consists of a cascaded 1×1 convolutional layer and an upsampling layer. The 1×1 convolutional layer is used to further extract and reduce the dimensionality of the features, and the upsampling layer is used to restore the IASPP input feature size. The output is calculated sequentially. and F c1 F c2 F c3 F c4 and F c5 By performing a concatenation, we obtain The input consists of a concatenated 3×3 convolutional layer and a 1×1 convolutional layer with a stride of 1. The output is calculated sequentially. and

[0054] Fl3rsb23 After the calculation process of the IASPP module and F l1rsb3 The calculation process for the input is consistent, and the last 1×1 convolutional layer also controls the number of output feature channels to be 256, F l3rsb23 As input, calculate the output.

[0055] (3) The IMSFF module adds an IASPP module after the first and third layers of IR-Net. Each IASPP module uses a 1×1 convolution to adjust the number of channels in the output feature map to n. Then, the feature map output from the third layer via the IASPP module is upsampled by 4x bilinear interpolation. Simultaneously, the output feature map from the first layer is processed by a 1×1 convolution to adjust the number of channels to n. The three feature maps are then fused side-by-side and finally processed by a 3×3 convolution to eliminate aliasing. The calculation process is as follows:

[0056] F l1rsb3 The outputs are processed through the IASPP module and a 1×1 convolutional layer, respectively. and F l3rsb23 The output is sequentially processed through the IASPP module and a 4x bilinear interpolation upsampling layer. and

[0057] F l1rsb3 F l1conv and F l3conv The feature map is obtained by juxtaposing the features.

[0058] F cc After eliminating aliasing and refining the features using a 3×3 convolutional layer with a stride of 1, the calculated...

[0059] For F c33 Perform 4x bilinear interpolation upsampling, then substitute the results into the softmax function to obtain the classification result. Since it is a binary classification task, the output is the probability of the foreground and background of the crack, so the number of channels is 2.

[0060] S104. Model Training: The model training process employs advanced hardware and meticulously designed training settings to ensure that the model can converge efficiently and achieve optimal performance.

[0061] (1) Hardware equipment:

[0062] Operating System: The operating system used during training is Ubuntu 2204LTS, which is known for its stability and widely supported software ecosystem, making it suitable for deep learning tasks.

[0063] CPU: The central processing unit used is an Intel Xeon E5-2680v4. This CPU has the advantages of multi-core and multi-thread, which can effectively support multi-task parallel processing in data preprocessing and model training.

[0064] GPU: The graphics processor used is the NVIDIA H200, a high-performance GPU with 141GB of video memory, capable of easily handling large-scale deep learning models. The H200 supports the CUDA 12.4 toolkit, providing efficient parallel computing capabilities and accelerating the model training process.

[0065] (2) Training settings:

[0066] Training environment configuration: Python version 3.9.13 was used, and the deep learning framework selected was PyTorch version 2.5.0, which provides excellent support for GPU acceleration and distributed training. The training code ran on an NVIDIA H200 GPU to fully utilize its computing power.

[0067] Optimizer selection: The Adam optimizer was used, an adaptive moment estimation optimization algorithm that automatically adjusts the learning rate, reducing the need for manual parameter tuning. The initial learning rate was set to 0.0001, and the weight decay was set to 0.0001. These parameters have been verified through multiple experiments to effectively prevent model overfitting.

[0068] Loss function: The cross-entropy loss function is used, which can effectively measure the difference between the predicted result and the true label. It is particularly suitable for handling class imbalance problems and ensures that the segmentation accuracy of the gap region and the background region is improved in a balanced way.

[0069] Learning rate adjustment strategy: A cosine annealing strategy is adopted. This strategy dynamically adjusts the learning rate, which enables the model to converge quickly in the early stage of training, and then finely adjusts it with a lower learning rate in the later stage of training, effectively avoiding getting trapped in local optima.

[0070] Training epochs: The model training is set to 100 epochs, and the model parameters are saved every 10 epochs to ensure that the model can be restored to its optimal state at any time during training and to prevent overfitting.

[0071] Batch size: Set to 8.

[0072] Model saving and evaluation: Model parameters are saved every 10 epochs. At the same time, the model is evaluated during each saving, and performance indicators such as mIoU, IoU and Dice coefficient are recorded for subsequent analysis and optimization.

[0073] S105. Model Evaluation: To verify the performance of the improved model, this experiment uses the following performance metrics for evaluation:

[0074] Mean Intersection over Union (mIoU): Used to measure the average segmentation accuracy of the model across all classes.

[0075] Intersection over Union (IoU): Used to measure the accuracy of a model in segmenting a particular class.

[0076] Dice coefficient: Used to measure the quality of a model's segmentation of a particular class.

[0077] The experimental results are as follows:

[0078] Model mIoU IoU (crack) Dice DeepLabV3+ 0.72 0.75 0.82 Improved Model 0.85 0.88 0.91

[0079] The experimental results show that the improved model outperforms the original model in terms of mIoU, IoU, and Dice coefficient.

[0080] The DeepLabV3+ model demonstrates that the improved model has better performance in the semantic segmentation task of tunnel face cracks.

[0081] Example 2

[0082] Combination Figure 4 As shown, this disclosure provides a tunnel face crack identification device 300 based on DeepLabV3+, including a processor 304 and a memory 301. Optionally, the device may further include a communication interface 302 and a bus 303. The processor 304, communication interface 302, and memory 301 can communicate with each other via the bus 303. The communication interface 302 can be used for information transmission. The processor 304 can call logical instructions in the memory 301 to execute the tunnel face crack identification method based on DeepLabV3+ described in the above embodiment.

[0083] Furthermore, the logic instructions in the aforementioned memory 301 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0084] The memory 301, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 304 executes functional applications and data processing by running the program instructions / modules stored in the memory 301, thereby implementing the tunnel face crack identification method based on DeepLabV3+ in the above embodiments.

[0085] The memory 301 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 301 may include high-speed random access memory and may also include non-volatile memory.

[0086] Example 3

[0087] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the tunnel face crack identification method based on DeepLabV3+ described in the above embodiments.

[0088] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0089] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0090] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for descriptive purposes only and is not intended to limit the scope of protection. As used in the description herein, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used herein means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0092] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

Claims

1. A method for identifying cracks at the tunnel face based on DeepLabV3+, characterized in that: Includes the following steps: S01. Obtain images of the tunnel face at the construction site and preprocess them; S02. Input the preprocessed tunnel face image into the tunnel face crack recognition model based on DeepLabV3+. The tunnel face crack recognition model uses the first three layers of the ResNet series model as the backbone network of DeepLabV3+. In each residual block of the second and third layers of the backbone network, the 3×3 convolution is replaced with two cascaded dilated convolutions, and the stride of the first convolution kernel of the first residual block of the second and third layers of the backbone network is set to 1. The backbone network completes the multi-scale feature extraction of the tunnel face image. The first and third layers of the backbone network are both connected to IASPP modules. The IASPP module adds a 3×3 convolutional layer before the 1×1 convolutional layer of the ASPP module to eliminate aliasing. The pooling of multi-scale features of the tunnel face image is completed through the two IASPP modules to obtain the deep feature map. The output features of the first layer of the backbone network are processed by the IASPP module and a 1×1 convolution to obtain feature F. l1iaspp F l1conv The output features of the third layer of the backbone network are obtained by upsampling through the IASPP module to obtain feature F. l3conv Feature F l1iaspp F l1conv F l3conv After splicing, feature F is obtained. cc Feature F cc After eliminating aliasing effects and refining the features using a 3×3 convolutional layer, feature F is obtained. c33 , for F c33 Upsampling is performed, and then the result F is calculated by substituting it into the softmax function. prediction .

2. The tunnel face crack identification method based on DeepLabV3+ according to claim 1, characterized in that: The first three layers of ResNet-101 are used as the backbone network of DeepLabV3+. The backbone network consists of the input layer, the first layer, the second layer and the third layer connected in series. The first layer consists of one max pooling layer and three residual blocks. Each residual block includes three concatenated 1×1 convolutional layers I, 3×3 convolutional layers, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer to the output of the three concatenated convolutional layers. The second layer consists of four concatenated residual blocks. Each residual block includes four concatenated 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer and the output of the four concatenated convolutional layers. The third layer consists of 23 residual blocks in series. Each residual block includes four cascaded 1×1 convolutional layers I, 3×3 dilated convolutional layers I, 3×3 dilated convolutional layers II, and 1×1 convolutional layers II, as well as a skip connection layer. The skip connection layer connects the input and output of this residual block. The output of this residual block is obtained by adding the output of the skip connection layer and the output of the four cascaded convolutional layers.

3. The tunnel face crack identification method based on DeepLabV3+ according to claim 2, characterized in that: The stride of each residual block in the first layer is 1 for the 1×1 convolutional layer I, the 3×3 convolutional layer, the 1×1 convolutional layer II, and the skip connection layer.

4. The tunnel face crack identification method based on DeepLabV3+ according to claim 2, characterized in that: In the second layer, the stride of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II in each residual block is 1. The rates of the 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4, respectively. The stride of the 1×1 convolutional layer I and the skip connection layer in the first residual block of the second layer is 2, and the stride of the 1×1 convolutional layer I and the skip connection layer in the last three residual blocks is 1.

5. The tunnel face crack identification method based on DeepLabV3+ according to claim 2, characterized in that: In the third layer, the stride of the 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 1×1 convolutional layer II in each residual block is 1. The rates of the 3×3 dilated convolutional layer I and 3×3 dilated convolutional layer II are 2 and 4, respectively. The stride of the 1×1 convolutional layer I and the skip connection layer in the first residual block of the third layer is 2, and the stride of the 1×1 convolutional layer I and the skip connection layer in the last 22 residual blocks is 1.

6. The method for identifying tunnel face cracks based on DeepLabV3+ according to claim 1, characterized in that: The IASPP module includes parallel 1×1 convolutional layers, 3×3 dilated convolutional layers I, II, and III, and a global average pooling layer. After the global average pooling layer, a 1×1 convolutional layer and an upsampling layer are connected. The input passes through these parallel layers to obtain the feature F. c1 F c2 F c3 F c4 F map F map The feature F is obtained after a 1×1 convolutional layer and an upsampling layer following a global average pooling layer. c5 , will feature Fc 1 F c2 F c3 F c4 and F c5 The feature F is obtained by splicing. concat Feature F concat After passing through a series of 3×3 convolutional layers and 1×1 convolutional layers, the output of the IASPP module is obtained.

7. The tunnel face crack identification method based on DeepLabV3+ according to claim 6, characterized in that: The rates of 3×3 dilated convolutional layer I, 3×3 dilated convolutional layer II, and 3×3 dilated convolutional layer III are 6, 12, and 18, respectively.

8. The tunnel face crack identification method based on DeepLabV3+ according to claim 6, characterized in that: The preprocessing includes normalization, data augmentation, and label processing.

9. A tunnel face crack identification device based on DeepLabV3+, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute, when running the program instructions, the tunnel face crack identification method based on DeepLabV3+ as described in any one of claims 1 to 8.

10. A storage medium storing program instructions, characterized in that, When the program instructions are executed, they perform the tunnel face crack identification method based on DeepLabV3+ as described in any one of claims 1 to 8.

Images